Linear evaluation weight problems ("_IncompatibleKeys" and access issues)

Hello, thanks for all the amazing work you put into this. I tried downloading the pretrained weights for the 3 available ViT models and have encountered some issues:

1. ViT-S/16 gives me an “Access denied” message whenever I try downloading it.
2. For ViT-S/8 or ViT-B/16, their weights seem corrupted? Whenever I try loading them into eval_linear.py, I get a message listing a long list of missing and unexpected keys. Not exactly sure what is wrong here. The loss also starts at a quite high value and although it’s dropping off, I don’t think this is the intended behavior from a pretrained model.

Here’s the training output (not the log) for ViT-S/8 TrainingOutput.txt

Thank you again for your work.

[mathildecaron31]

hi @kikacaty can you do a fresh clone and run again ? The problem with vit_base is due to a discrepency in the way I concatenated the features (related to https://github.com/facebookresearch/dino/issues/134) but that should be fixed in this commit https://github.com/facebookresearch/dino/commit/d2f3156bb34c32ae54312de7fe4e31580f12ff7f

[mathildecaron31]

@KnockerPulsar

I have tried to improve the code so you can easily load the pretrained linear weights and check their performance (situation b. in my post above). For example you can run:

python eval_linear.py --evaluate --arch vit_small --patch_size 16 --data_path /path/to/imagenet/train

python eval_linear.py --evaluate --arch vit_small --patch_size 8 --data_path /path/to/imagenet/train

python eval_linear.py --evaluate --arch vit_base --patch_size 16 --n_last_blocks 1 --avgpool_patchtokens true --data_path /path/to/imagenet/train

python eval_linear.py --evaluate --arch vit_base --patch_size 8 --n_last_blocks 1 --avgpool_patchtokens true --data_path /path/to/imagenet/train

python eval_linear.py --evaluate --arch resnet50 --data_path /path/to/imagenet/train

https://github.com/facebookresearch/dino/commit/d2f3156bb34c32ae54312de7fe4e31580f12ff7f

Hope that helps.